Retargeting and Respecializing GPU Workloads for Performance Portability